Note: This task is essentially identical to the extended data task of the 2002 evaluation. The data, tables and trials all remain the same. This task does include some new auxiliary information.

Auxiliary Information:

Available after the official evaluation:

SRI has made available word and phone alignments for the SRE03 evaluation data. The tar file contains a README file describing the distribution.

Available for use during the official evaluation:

1. Automatically generated word transcripts from a realtime recognizer (the same transcriptions as provided in 2002)
   
   

   These ASR transcripts were generated at NIST, using a version of BBN's Byblos system. This was a near real time system which was NOT optimized for this task, and therefore the word error rate is expected to be much higher than it was for the transcripts provided for last year's dry run evaluation.

2. Automatically generated phone level transcriptions for five different languages (provided by R523)
   
   

   The transcripts are from phone recognizers for the following five languages: English, German, Japanese, Mandarin and Spanish.
   
   -rw-rw-r-- 175134720 PHONES_SRE03_edt.tar

3. Pitch track estimates (provided by SRI)
   
   

   The pitch track estimates are on a 8-CDROM set. They were mailed with the evaluation data to the sites that have previously requested them.
   
   Send e-mail if you require a copy of this data. NOTE, this data is available to participants in the 2003 extend data task only!

4. Base GMM-UBM scores (provided by MIT-LL)
   
   

   The GMM-UBM scores are from MIT-LL's 2002 submission of a base system without further fusion using other information.
   
   -rw-rw-r-- 801104 GMM_UBM_SRE03_edt.tar.gz
   -rw--r--r-- 2230 missing_167.txt.gz

5. Automatically generated handset type labels (provided by MIT-LL)

   The handset labels and gender identifier were automatically produced by MIT-LL using a different handset labeler than was used for previous NIST evaluations. The format of each file is:
   <SWB_FILE> <CHANNEL A|B> <HANDSET CARB|ELEC> <GENDER M|F>
   
   -rw-rw-r-- 41331 Handsets_gender_SRE03_edt.tar.gz
   

6. Language Model probabilities (provided by AFRL/HECA)
   
   

   The LM probabilities are given for each of the splits defined in the control file. AFRL/HECA has provided a detailed readme.txt file, located in the release.
   
   -rw-rw-r-- 1309980 LM_SRE03_edt.tar.gz

7. Speech Activity Detection Labels (provided by MIT-LL)
   
   

   The SAD file contains a readme.txt to describe the file format. This was provided to accompany the Pitch track estimates.
   
   -rw-rw-r-- 7178240 SAD_SRE03_edt.tar
   

Tables and Control File (same as in the 2002 evaluation)

Speaker Conversation Table

Evaluation Control File [ NEW March 21st, 2002 ]

Development ASR Transcripts

• ~24MB Switchboard 1 ASR transcripts (4876 sides)

[ SRE2003 ]